Distributed BMS battery pack health assessment method based on federated learning

Abstract

The invention discloses a distributed BMS battery pack health assessment method based on federated learning. According to the method, the federal learning architecture is introduced, so that each edge node can finish data acquisition, feature extraction and model training locally, only model parameters instead of original data are uploaded, and the leakage risk caused by concentrated transmission of a large amount of sensitive data is fundamentally avoided. And meanwhile, the coordination server adopts a weighted aggregation strategy to allocate weights according to node sample sizes, so that knowledge contribution of large sample nodes is fully embodied, and global model bias caused by node number difference is avoided. According to the mode of information fusion instead of data fusion, data privacy of all nodes is protected, cross-node knowledge sharing is achieved, and the system can fully utilize multi-source heterogeneous data in the distributed BMS to conduct collaborative modeling while the strict privacy protection requirement is met.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery management systems, and specifically provides a distributed BMS battery pack health assessment method based on federated learning. Background Art

[0002] With the rapid development of applications such as electric vehicles and energy storage systems, lithium-ion battery packs are expanding in size and becoming increasingly complex. Traditional centralized battery management systems (BMS) are no longer able to meet the real-time, reliability, and scalability requirements. To address the communication pressures and computing bottlenecks associated with the scale of battery systems, distributed BMSs have become a hot topic in research and industry development. However, in a distributed BMS architecture, battery cluster nodes operate in a decentralized manner, resulting in significant data heterogeneity.

[0003] However, traditional centralized modeling health assessment methods face multiple challenges: first, there is a lack of a unified modeling mechanism, which makes model migration and coordination difficult; second, the original data between nodes cannot be directly aggregated, which is limited by communication bandwidth and poses serious data privacy and security risks; third, in a multi-working environment, node model training is prone to overfitting, affecting the overall prediction accuracy and stability.

[0004] In recent years, federated learning, a new distributed machine learning framework, has emerged as a promising approach for collaborative multi-device model training without sharing raw data. It has found promising applications in healthcare, finance, and edge computing. Its introduction into distributed BMS health assessment systems promises to overcome the bottlenecks of traditional approaches in data fusion, model generalization, and privacy protection. Therefore, a distributed BMS health assessment method that incorporates federated learning mechanisms is urgently needed to enhance the overall intelligence and operational capabilities of the system. Summary of the Invention

[0005] The purpose of the present invention is to provide a distributed BMS battery pack health assessment method based on federated learning in order to solve the above-mentioned problems.

[0006] The technical solution adopted by the present invention is as follows: a distributed BMS battery pack health assessment method based on federated learning, the method comprising the following steps:

[0007] S1: Perform system deployment and federation structure initialization, organizing multiple physically separated but functionally collaborative BMS subsystem nodes, referred to as "edge nodes", into a distributed learning network;

[0008] S2: edge node data collection and training sample construction. After the system deployment is completed, each edge node begins to perform its local data collection task. The present invention proposes to segment the continuous operating condition data in a "sliding window" manner to generate structured and standardized modeling samples.

[0009] S3: Build and train a local health assessment model. During the local training phase, the present invention deploys a health assessment model on each edge node to learn the mapping relationship between feature vectors and SOH. The model adopts a feedforward neural network structure, which has the advantages of simple structure, efficient computation, and easy training and migration, making it particularly suitable for deployment on resource-constrained embedded platforms.

[0010] S4: Perform model upload and federation aggregation mechanism. After training, the present invention only uploads local model parameters instead of original data to the coordination server under the premise of ensuring data privacy.

[0011] S5: Introduce differential privacy mechanism to protect uploaded parameters;

[0012] S6: Perform personalized fine-tuning and model fusion. The fine-tuning method is to add the original model θ (t+1) Based on this, 1 to 3 rounds of local training are performed to form personalized parameters θ i personal ;

[0013] S7: Output health score and status classification. After all model training and fine-tuning are completed, each node uses its final model to infer the current data segment and output the predicted health score.

[0014] In a preferred embodiment, in step S1, each edge node in the system is abstracted as an intelligent agent with local data collection, feature extraction, model training and online prediction capabilities. At the same time, a "federal coordination server" located on the master terminal or cloud is set up to collect model parameters uploaded by all child nodes and complete the aggregate calculation and feedback update of the global model.

[0015] During the initialization phase, the coordination server presets a unified structural framework for the health assessment model and sends the initial model parameters to each node. This model structure must meet the following two basic requirements: first, it must be lightweight enough to run in an embedded environment or a low-computing-power processor; second, it must have certain nonlinear modeling capabilities to fully explore the complex relationship between battery status characteristics and health.

[0016] In a preferred embodiment, in step S1, the initial model adopts a fully connected neural network, including an input layer, one or two hidden layers (ReLU activation) and an output layer (Sigmoid activation); the output is a continuous variable between 0 and 1, representing the health score of the current battery cluster (SOH estimate); the goal of this stage is to complete the full system federated training preparation work of "unified communication structure establishment + model skeleton synchronization + training mechanism preparation".

[0017] In a preferred embodiment, in step S2, the data collection content includes but is not limited to state variables such as cell voltage, cell temperature, battery cluster current (charge and discharge current), SOC trajectory (state of charge), timestamp and cycle count; the node divides the above data into time sliding windows W, and each window is used as an independent sample segment; statistics and engineering feature extraction are performed on each segment, and common indicators such as voltage range ΔV = max (V) - min (V) (measurement of voltage consistency), average temperature (used to reflect thermal stability), SOC slip (revealing the dynamic rate of charge change), capacity increment peak (used to identify aging characteristics) and current mean and change rate, etc.;

[0018] The eigenvectors are obtained by concatenating x i j ∈R d , as the model input; and the corresponding health label y i j ∈[0,1] can come from capacity calibration experiments, SOH estimation models, or redundant sensor judgment results; ultimately, each node constructs a local training sample set:

[0019]

[0020] where n i Represents the number of samples, and j represents the sample number; this process realizes the automated transformation from "raw multidimensional time series data" to "standardized modeling input", ensuring that each node has the foundation for local modeling capabilities.

[0021] In a preferred embodiment, in step S3, the model adopts a feedforward neural network structure, which has the advantages of simple structure, high computational efficiency, ease of training and migration, and is particularly suitable for deployment on resource-constrained embedded platforms;

[0022] The input is the feature vector x, and the output is the health score Where θ is the model parameter set; during the training process, the mean square error (MSE) is used as the main loss term, and the L2 regularization term is introduced to suppress overfitting, forming the objective function:

[0023]

[0024] The first term measures the prediction error, and the second term prevents the weight from being infinitely amplified. The regularization coefficient λ is usually set to 1×10 -4 ~1×10 -3 .

[0025] In a preferred embodiment, in step S3, the training optimization uses an adaptive gradient descent algorithm such as the Adam optimizer or RMSprop to improve the convergence speed and stability; each round of local training includes multiple epochs, and the mini-batch technology is used for data partitioning, with a typical batch size of 32 to 64; after the training is completed, each node will retain the model parameters θi obtained from the current training for subsequent federated synchronization.

[0026] In a preferred embodiment, in step S4, the model parameters θi uploaded by each node can be understood as the knowledge representation in the health assessment model structure learned by the current node; after receiving the parameters uploaded by all nodes, the coordination server performs a weighted aggregation operation to form a new global model parameter θ (t+1) :

[0027]

[0028] This aggregation strategy (called FedAvg) assigns weights based on the node sample size, ensuring that nodes with large samples have greater influence without causing global model bias due to differences in the number of nodes; this process completes "information fusion rather than data fusion", fully retaining the knowledge contribution of each node while avoiding the risk of data leakage; after the global model is generated, the coordination server will synchronously send it down as the initial model for the next round of node training, starting a new round of federated iteration.

[0029] In a preferred embodiment, in step S5, a specific method is to inject random noise that obeys Gaussian distribution before uploading the parameters:

[0030]

[0031] The noise intensity σ can be automatically adjusted according to the preset privacy budget (∈, δ) to achieve differential privacy protection. In addition, to improve the robustness of the model under non-independent and identically distributed (Non-IID) data distribution conditions, this paper also introduces a proximal regularization term, which adds a penalty term for global parameters to the local loss function:

[0032]

[0033] It can effectively limit the node model from overfitting locally and deviating from the global direction.

[0034] In a preferred embodiment, in step S6, after the global model is aggregated, the present invention allows each node to perform "personalized fine-tuning" to adapt to local characteristics; the fine-tuning method is to add a new node to the original model θ (t+1) Based on this, 1 to 3 rounds of local training are performed to form personalized parameters θ i personal ; When making the final prediction, the model can use a weighted fusion approach to comprehensively consider individual and global contributions:

[0035]

[0036] The weight λ can be dynamically adjusted according to the node model stability or error history to ensure the accuracy and consistency of the final prediction results.

[0037] In a preferred embodiment, in step S7, the node can upload the score back to the coordination server for system-level health assessment, degradation trend analysis and heat map display, providing a decision-making basis for operation and maintenance personnel. At the same time, operation and maintenance personnel can adjust the threshold of the health level score according to actual conditions, which is more flexible.

[0038] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0039] 1. In the present invention, by introducing a federated learning architecture, each edge node can complete data collection, feature extraction, and model training locally, and only upload model parameters instead of original data, fundamentally avoiding the risk of leakage caused by the centralized transmission of large amounts of sensitive data. At the same time, the coordination server adopts a weighted aggregation strategy to assign weights according to the sample size of the nodes, ensuring that the knowledge contribution of large sample nodes is fully reflected and that global model bias is not caused by differences in the number of nodes. This "information fusion rather than data fusion" approach not only protects the data privacy of each node, but also realizes cross-node knowledge sharing, allowing the system to meet strict privacy protection requirements while fully utilizing multi-source heterogeneous data in the distributed BMS for collaborative modeling.

[0040] 2、In the present application, the evaluation accuracy and system stability are significantly improved by using multiple technical means. The introduced Proximal regular term can effectively limit the local overfitting of the node model, and avoid the deviation of the model from the global direction caused by the non-independent and identically distributed data distribution; the differential privacy mechanism injects Gaussian noise to provide reliable privacy protection for the uploaded parameters without affecting the model performance. The personalized fine-tuning function allows each node to perform a small amount of local training based on the global model, so that the model can better adapt to the data differences caused by different battery types, brands, installation locations and temperature environments. The health score uses a continuous variable output of 0-1, and gives operation and maintenance suggestions combined with grading standards, which not only provides intuitive judgment basis for operation and maintenance personnel, but also supports flexible adjustment of threshold according to actual situation, so that the evaluation result can directly guide the maintenance decision of the battery system, effectively improving the intelligent level and operation and maintenance efficiency of the large battery system health management. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The flow principle diagram of the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical scheme and advantages of the present application more clear and explicit, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0043] Embodiment:

[0044] Reference Figure 1 ,

[0045] A battery pack health evaluation method fusing a federal learning mechanism. The method realizes efficient cooperation and knowledge sharing of model parameters among multiple nodes by introducing a federal learning architecture in a distributed BMS system, completes distributed training and dynamic updating of a global health evaluation model on the premise of ensuring that private data does not leave the local, and considers evaluation accuracy, system real-time performance and deployment feasibility. The overall scheme is suitable for large battery system health management tasks facing complex operating conditions. The technical scheme will be described in detail below, and the flow chart of the method is shown in Figure 1 .

[0046] Step 1: System deployment and federal structure initialization The reality of multiple battery cluster nodes independently running in the BMS system, the traditional centralized modeling method has been unable to adapt to the current needs of battery systems in terms of heterogeneity, privacy and real-time performance. Therefore, the federal learning idea is introduced, and multiple physically separated but functionally cooperative BMS subsystem nodes (hereinafter referred to as "edge nodes") are organized into a distributed learning network.

[0047] In the system, each edge node is abstracted as an intelligent agent (Agent) with local data collection, feature extraction, model training and online prediction capabilities. At the same time, a "federal coordination server" located in the host or cloud is set up to collect all the model parameters uploaded by the sub-nodes and complete the aggregation calculation and return update of the global model.

[0048] In the initialization phase, the coordination server presets the unified structure framework of the health assessment model, and delivers the initial model parameters to each node. This model structure needs to meet the following two basic requirements: first, it is light enough to run in embedded environment or low-power processor; second, it has certain nonlinear modeling capability to fully explore the complex relationship between battery state characteristics and health degree.

[0049] The initial model uses a fully connected neural network, including an input layer, one or two hidden layers (ReLU activation), and an output layer (Sigmoid activation). The output is a continuous variable between 0 and 1, representing the health score (SOH estimate) of the current battery cluster.

[0050] The main goal of this phase is to complete the "unified communication structure establishment + model skeleton synchronization + training mechanism preparation" of the whole system federal training preparation work.

[0051] Step 2: Edge node data collection and training sample construction

[0052] After the system deployment is completed, each edge node starts to perform its local data collection task. This invention proposes to segment the continuous running condition data in a "sliding window" manner, thereby generating structured and standardized modeling samples.

[0053] The data collection content includes but is not limited to single cell voltage, single cell temperature, battery cluster current (charge and discharge current), SOC trajectory (state of charge), timestamp and cycle count, etc. The node divides the above data into time sliding windows W, and each window is regarded as an independent sample segment. Statistical and engineering feature extraction is performed on each segment. Common indicators include: voltage range ΔV = max(V) - min(V) (measuring voltage consistency), average temperature (reflecting thermal stability), SOC slip (revealing the rate of charge dynamic change), capacity increment peak (used to identify aging characteristics), and current mean and change rate, etc.

[0054] The feature vector is obtained by concatenation x i j ∈R d , as the model input. And the corresponding health label y i j∈[0,1] can be derived from capacity calibration experiments, SOH estimation models, or redundant sensor judgment results. Finally, each node constructs a local training sample set:

[0055]

[0056] where n i This process automatically transforms "raw multidimensional time series data" into "standardized modeling input," ensuring that each node has the foundation for local modeling capabilities.

[0057] Step 3: Local health assessment model construction and training

[0058] During the local training phase, the present invention deploys a health assessment model at each edge node to learn the mapping relationship between feature vectors and SOH. This model uses a feedforward neural network architecture, which offers advantages such as simplicity, computational efficiency, and ease of training and migration, making it particularly suitable for deployment on resource-constrained embedded platforms.

[0059] The input is the feature vector x, and the output is the health score Where θ is the model parameter set. During the training process, the mean square error (MSE) is used as the main loss term, and the L2 regularization term is introduced to suppress overfitting, forming the objective function:

[0060]

[0061] The first term measures the prediction error, and the second term prevents the weight from being infinitely amplified. The regularization coefficient λ is usually set to 1×10 -4 ~1×10 -3 .

[0062] Training optimization uses adaptive gradient descent algorithms such as the Adam optimizer or RMSprop to improve convergence speed and stability. Each round of local training consists of multiple epochs, and data is partitioned using mini-batch technology, with a typical batch size of 32 to 64. After training, each node retains the model parameters θi obtained from the current training for subsequent federated synchronization.

[0063] Step 4: Model upload and federation aggregation mechanism

[0064] After training, the present invention only uploads local model parameters rather than raw data to the coordination server while ensuring data privacy. The model parameters θi uploaded by each node can be understood as the knowledge representation of the health assessment model structure learned by the current node. After receiving the parameters uploaded by all nodes, the coordination server performs a weighted aggregation operation to form a new global model parameter θ (t+1) :

[0065]

[0066] This aggregation strategy (called FedAvg) assigns weights based on node sample size, ensuring that nodes with large sample sizes have greater influence without biasing the global model due to differences in node numbers. This process achieves "information fusion rather than data fusion," fully preserving the knowledge contribution of each node while minimizing the risk of data leakage. Once the global model is generated, the coordination server simultaneously distributes it as the initial model for the next round of node training, initiating a new round of federated iterations.

[0067] Step 5: Privacy protection mechanism and model stability improvement

[0068] In terms of system security, this invention protects uploaded parameters by introducing a differential privacy mechanism. Specifically, this mechanism injects random noise that follows a Gaussian distribution before uploading the parameters:

[0069]

[0070] The noise intensity σ can be automatically adjusted according to the preset privacy budget (∈, δ) to achieve differential privacy. In addition, to improve the robustness of the model under non-independent and identically distributed (Non-IID) data distribution conditions, this paper also introduces a proximal regularization term, which adds a penalty term for global parameters to the local loss function:

[0071]

[0072] It can effectively limit the node model from overfitting locally and deviating from the global direction.

[0073] Step 6: Personalized fine-tuning and model fusion

[0074] Since there are differences in battery type, brand, installation location, temperature environment, etc. in actual working conditions, even if the model structure is consistent, the data distribution between different nodes is extremely heterogeneous. Therefore, after the global model is aggregated, the present invention allows each node to perform "personalized fine-tuning" to adapt to local characteristics. The fine-tuning method is to add a new parameter to the original model θ (t+1) Based on this, 1 to 3 rounds of local training are performed to form personalized parameters θ i personal In the final prediction, the model can use a weighted fusion approach to comprehensively consider individual and global contributions:

[0075]

[0076] The weight λ can be dynamically adjusted according to the node model stability or error history to ensure the accuracy and consistency of the final prediction results.

[0077] Step 7: Health score output and status classification

[0078] After all model training and fine-tuning are completed, each node uses its final model to infer the current data segment and output a predicted health score. In order to enhance the readability and controllability of the system on the operation and maintenance side, the present invention sets the health level classification standard:

[0079]

[0080]

[0081] Nodes can upload scores back to the coordination server for system-level health assessment, degradation trend analysis, and heat map display, providing decision-making basis for operation and maintenance personnel. At the same time, operation and maintenance personnel can adjust the threshold of the health level score according to actual conditions, which is more flexible.

[0082] From the above we can know:

[0083] In the present invention, by introducing a federated learning architecture, each edge node can complete data collection, feature extraction, and model training locally, and only upload model parameters instead of raw data, fundamentally avoiding the risk of leakage caused by the centralized transmission of large amounts of sensitive data. At the same time, the coordination server adopts a weighted aggregation strategy to assign weights according to the sample size of the nodes, ensuring that the knowledge contribution of large sample nodes is fully reflected and that global model bias is not caused by differences in the number of nodes. This "information fusion rather than data fusion" approach not only protects the data privacy of each node, but also realizes cross-node knowledge sharing, allowing the system to meet strict privacy protection requirements while fully utilizing multi-source heterogeneous data in the distributed BMS for collaborative modeling.

[0084] In the present invention, the evaluation accuracy and system stability are significantly improved through a number of technical means. The introduced Proximal regularization term can effectively limit the local overfitting of the node model and avoid the deviation of the model from the global direction due to the non-independent and identically distributed data distribution; the differential privacy mechanism injects Gaussian noise to provide reliable privacy protection for uploaded parameters without affecting the performance of the model. The personalized fine-tuning function allows each node to perform a small amount of local training based on the global model, so that the model can better adapt to the data differences caused by different battery types, brands, installation locations and temperature environments. The health score is output as a continuous variable of 0 to 1, and operation and maintenance recommendations are given in combination with the grading standard. It not only provides an intuitive judgment basis for operation and maintenance personnel, but also supports flexible adjustment of thresholds according to actual conditions, so that the evaluation results can directly guide the maintenance decisions of the battery system, effectively improving the intelligence level and operation and maintenance efficiency of the health management of large-scale battery systems.

[0085] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0086] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A distributed BMS battery pack health assessment method based on federated learning, characterized by: The method comprises the following steps: S1: Perform system deployment and federation structure initialization, organizing multiple physically separated but functionally collaborative BMS subsystem nodes, referred to as "edge nodes", into a distributed learning network; S2: Edge node data collection and training sample construction. After the system deployment is completed, each edge node begins to perform its local data collection task. The present invention proposes to segment the continuous operating condition data in a "sliding window" manner to generate structured and standardized modeling samples. S3: Build and train a local health assessment model. During the local training phase, the present invention deploys a health assessment model on each edge node to learn the mapping relationship between feature vectors and SOH. The model adopts a feedforward neural network structure, which has the advantages of simple structure, efficient computation, and easy training and migration, making it particularly suitable for deployment on resource-constrained embedded platforms. S4: Perform model upload and federation aggregation mechanism. After training, the present invention only uploads local model parameters instead of original data to the coordination server under the premise of ensuring data privacy. S5: Introduce differential privacy mechanism to protect uploaded parameters; S6: Perform personalized fine-tuning and model fusion. The fine-tuning method is to add the original model θ (t+1) Based on this, 1 to 3 rounds of local training are conducted to form personalized parameters S7: Output health score and status classification. After all model training and fine-tuning are completed, each node uses its final model to infer the current data segment and output the predicted health score.

2. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S1, each edge node in the system is abstracted as an intelligent agent with local data collection, feature extraction, model training, and online prediction capabilities. At the same time, a "federal coordination server" located on the master terminal or cloud is set up to collect model parameters uploaded by all child nodes and complete the aggregate calculation and feedback update of the global model. During the initialization phase, the coordination server presets a unified structural framework for the health assessment model and sends the initial model parameters to each node. This model structure must meet the following two basic requirements: first, it must be lightweight enough to run in an embedded environment or a low-computing-power processor; second, it must have certain nonlinear modeling capabilities to fully explore the complex relationship between battery status characteristics and health.

3. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S1, the initial model uses a fully connected neural network, including an input layer, one or two hidden layers (ReLU activation), and an output layer (Sigmoid activation); the output is a continuous variable between 0 and 1, representing the health score of the current battery cluster (SOH estimate); the goal of this stage is to complete the full system federated training preparation work of "unified communication structure establishment + model skeleton synchronization + training mechanism preparation".

4. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S2, the data collection content includes but is not limited to single cell voltage, single cell temperature, battery cluster current (charge and discharge current), SOC trajectory (state of charge), timestamp and cycle count and other state variables; the node divides the above data into time sliding windows W, and each window is used as an independent sample segment; statistics and engineering feature extraction are performed on each segment, and common indicators such as voltage range ΔV = max (V) - min (V) (measurement of voltage consistency), average temperature (used to reflect thermal stability), SOC slip (revealing the dynamic rate of charge change), capacity increment peak (used to identify aging characteristics) and current mean and change rate, etc.; The eigenvectors are obtained by concatenating x i j ∈R d , as the model input; and the corresponding health label y i j ∈[0,1] can come from capacity calibration experiments, SOH estimation models, or redundant sensor judgment results; ultimately, each node constructs a local training sample set: where n i Represents the number of samples, and j represents the sample number; this process realizes the automated conversion from "raw multidimensional time series data" to "standardized modeling input", ensuring that each node has the foundation for local modeling capabilities.

5. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S3, the model adopts a feedforward neural network structure, which has the advantages of simple structure, high computational efficiency, ease of training and migration, and is particularly suitable for deployment on resource-constrained embedded platforms; The input is the feature vector x, and the output is the health score Where θ is the model parameter set; during the training process, the mean square error (MSE) is used as the main loss term, and the L2 regularization term is introduced to suppress overfitting, forming the objective function: The first term measures the prediction error, and the second term prevents the weight from being infinitely amplified. The regularization coefficient λ is usually set to 1×10 -4 ~1×10 -3 .

6. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S3, training optimization uses an adaptive gradient descent algorithm such as the Adam optimizer or RMSprop to improve convergence speed and stability. Each round of local training includes multiple epochs, and uses mini-batch technology for data partitioning, with a typical batch size of 32 to 64. After training is completed, each node will retain the model parameters θi obtained from the current training for subsequent federated synchronization.

7. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S4, the model parameters θi uploaded by each node can be understood as the knowledge representation in the health assessment model structure learned by the current node; after receiving the parameters uploaded by all nodes, the coordination server performs a weighted aggregation operation to form a new global model parameter θ (t+1) : This aggregation strategy (called FedAvg) assigns weights based on the sample size of nodes, ensuring that nodes with large samples have greater influence without causing global model bias due to differences in the number of nodes. This process completes "information fusion rather than data fusion", fully retaining the knowledge contribution of each node while avoiding the risk of data leakage. After the global model is generated, the coordination server will synchronously distribute it as the initial model for the next round of node training, starting a new round of federated iteration.

8. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S5, a specific method is to inject random noise that obeys Gaussian distribution before uploading the parameters: The noise intensity σ can be automatically adjusted according to the preset privacy budget (∈, δ) to achieve differential privacy protection. In addition, to improve the robustness of the model under non-independent and identically distributed (Non-IID) data distribution conditions, this paper also introduces a proximal regularization term, which adds a penalty term for global parameters to the local loss function: It can effectively limit the node model from overfitting locally and deviating from the global direction.

9. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S6, after the global model is aggregated, the present invention allows each node to perform "personalized fine-tuning" to adapt to local characteristics; The fine-tuning method is to use the original model θ (t+1) Based on this, 1 to 3 rounds of local training are conducted to form personalized parameters When making the final prediction, the model can use a weighted fusion approach to comprehensively consider individual and global contributions: The weight λ can be dynamically adjusted according to the node model stability or error history to ensure the accuracy and consistency of the final prediction results.

10. The distributed BMS battery pack health assessment method based on federated learning according to claim 1, characterized in that: In step S7, the node can upload the score back to the coordination server for system-level health assessment, degradation trend analysis and heat map display, providing a decision-making basis for operation and maintenance personnel. At the same time, operation and maintenance personnel can adjust the threshold of the health level score according to actual conditions, which is more flexible.

Images